CN110472500A - A kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat - Google Patents
A kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat Download PDFInfo
- Publication number
- CN110472500A CN110472500A CN201910617212.7A CN201910617212A CN110472500A CN 110472500 A CN110472500 A CN 110472500A CN 201910617212 A CN201910617212 A CN 201910617212A CN 110472500 A CN110472500 A CN 110472500A
- Authority
- CN
- China
- Prior art keywords
- network
- high speed
- unmanned boat
- detection
- water surface
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Image Analysis (AREA)
Abstract
The water surface sensation target fast algorithm of detecting based on high speed unmanned boat that the invention discloses a kind of can in real time, robustly detect multiclass specific objective waterborne using the present invention.To the real-time and robustness demand of vision-based detection when the present invention is for high speed unmanned boat autonomous navigation, based on MobileNet convolutional layer design lightweight master network with rapidly extracting full figure feature, the output of multi-characteristic in SSD thought design detection sub-network network fusion master network is then based on as a result, by regression forecasting location parameter while non-maxima suppression redundant results to realize quick multiple scale detecting.It is that algorithm is realized and is verified on embedded gpu NVIDIA Jetson TX2 hardware platform the result shows that, it being capable of real-time detection multiclass specific objective waterborne using the present invention, has the characteristics that strong robustness, multiple dimensioned, the detection time of single frame video can control within 50ms.
Description
Technical field
The present invention relates to unmanned boat technical field of vision detection, and in particular to a kind of water surface vision based on high speed unmanned boat
Algorithm of Fast Object Detection.
Background technique
Unmanned surface vehicle is a kind of pier for having autonomous path planning and independent navigation ability, in military and civilian
Field has a wide range of applications.In practical application, unmanned boat autonomous navigation and completes task in high dynamic environment, it is desirable that
Have perception environment, discovery potential threat and the ability for executing Rational Path, need to generally be equipped with ultrasonic probe, visual sensing
The sensors such as device, millimetre-wave radar, X-band radar, laser range finder obtain environment and obstacle information.Ultrasonic probe and millimeter
Wave radar coverage is smaller, and laser range finder detects wide coverage, but condition is harsher, and X-band radar closely detects
Effect is bad, and the sensing range of visual sensor is more appropriate.In addition the information content that the sensors such as radar obtain is less, can
It carries out obstacle detection but can not classify, and obstacle classification is classified dangerous situation, situation judgement important role.Depending on
Feel that sensor has high-resolution, low cost, advantage easy to install and wide the visual field, the image of acquisition is handled in steady picture and defogging
Picture quality is higher afterwards, so it is the most universal to carry out unmanned boat detection of obstacles using vision algorithm at present.
Scene is complicated in the image of unmanned surface vehicle camera acquisition, often has multiple barriers, also frequently encounters bloom
According to, sea surface reflection, the spray interference situations such as, therefore realize detection of obstacles it is extremely difficult.S.Fefilatyev etc. is detected first
Then sea horizon searches for potential obstacle target in the region under sea horizon, determine barrier by the testing result of successive frame
Hinder object.Wang et al. and X.Mou et al. also use similar thinking, first detect sea horizon, then pass through conspicuousness or the overall situation
The methods of sparse potential barrier of detection.Matej Kristan et al. assumes that marine environment picture is divided into sky, land
The semantic region parallel and completely different with sea fog or the ship being docked near sea horizon, three, water body, by applying weak structure
Constraint derives model parameter, the estimation of segmentation mask and optimization algorithm using markov random file frame.With engineering
The extensive use for practising algorithm, has researcher that the method for machine learning is introduced into unmanned surface vehicle obstacle target detection field.
C.Li et al. by objectness method obtain potential target, and by merge with conspicuousness method obtain finally as a result,
Reach preferable detection accuracy.
The desin speed of current unmanned surface vehicle most in the world is chiefly used in battle reconnaissance, island more than 40 sections
In the complex water areas environment such as reef patrol, therefore there is higher requirement: 1. strong robustnesses to detection of obstacles identification, is carried on the back in extra large day
Can there be preferable detection effect when scape, offshore situation, illumination variation, multiple target scene；2. multiple scale detecting, nobody
Ship working environment is complicated, is likely encountered a variety of ship targets of different sizes；3. strong real-time, the higher situation of the unmanned boat speed of a ship or plane
Under require algorithm that can quickly detect；It is subsequent dangerous point 4. identification classification, algorithm can identify different targets and classification
Grade and situation judgement provide information.Existing algorithm or robustness are insufficient or real-time is poor or application conditions are harsh,
The application scenarios of high speed unmanned boat can not be perfectly suitable for.And the deep learning algorithm developed rapidly in recent years has robust
Property strong, multiple scale detecting identification the advantages of, but there are also limitations: the calculating higher cost of 1. network this body structure；2. lacking
High performance embedded gpu board is supported.
Summary of the invention
In view of this, the present invention provides a kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat, needle
This problem is identified to high speed surface unmanned boat detection of obstacles, algorithm model, Ke Yi are analyzed and established in conjunction with actual conditions
It is single while keeping 82.1% average detected precision to the obstacle targets such as different size of ship and motor boat under more scenes
The detection time of frame video is less than 50ms, better than current same type scheme.
To achieve the above object, the invention provides the following technical scheme:
A kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat, total algorithm model can be divided into two
Point, first part based on MobileNet convolutional layer designs lightweight master network to generate the characteristic patterns of multiple and different scales, and second
It is based partially on SSD thought and has built detection sub-network network, merge each layer characteristic pattern and final output.
One, convolutional layer is separated with reference to the conceptual design depth of separation convolution in lightweight master network, by Standard convolution layer
Resolve into one convolution of a depth convolution sum: each convolution kernel is applied to each channel by depth convolution, and puts convolution
For combining the output of channel convolution, greatly reducing the parameter amount of network model and calculating cost.
Depth separates the computing cost ratio of convolutional layer comparison with standard convolutional layer are as follows:
In formula: DFFor the width and height of input feature vector figure；M is the quantity of input channel；DKFor Standard convolution layer and depth
The convolution kernel size of convolution；N is the quantity of convolution kernel.
The introducing that depth separates convolutional layer also greatly reduces the parameter amount of network model, with currently a popular network model
Parameter amount be compared as follows shown in table:
Two, lightweight master network removes the last average pond layer of MobileNet, full articulamentum and Softmax layers, and
8 convolutional layers are added behind network to improve ability in feature extraction, each layer structural parameters of addition are as shown in the table:
Three, detection sub-network network extracts that various sizes of characteristic pattern is multiple dimensioned to realize as inputting, this 6 characteristic patterns it is big
Small is respectively 19 × 19,10 × 10,5 × 5,3 × 3,2 × 2,1 × 1, and generates one to each characteristic area of input feature vector figure
The default frame of serial different size, different proportion, size and ratio are related with corresponding characteristic layer, it is assumed that adopt when model inspection
With m layers of characteristic pattern, then the default frame ratio calculation formula of k-th of characteristic pattern is as follows:
[1, m] k ∈ in formula, wherein SmaxAnd SminRepresent default frame minimum and maximum ratio shared in character pair figure
Example, is respectively set to 0.2 and 0.9.
By providing different the ratio of width to height to the default frame on same characteristic layer to enhance the robustness to body form, fit
Detection case for different size of ship.Setting the ratio of width to height is aspect_ratio=(1,2,1/2,3,1/3), specific wide
It is determined with height by following formula:
Particularly, when aspect_ratio is 1, it is specified that default frame parameter are as follows:
Characteristic pattern makes it also can be corresponding different in the receptive field of picture with the difference of corresponding frame size, forms more rulers
The detection of degree.
Four, target classification and position are returned simultaneously when the training of detection sub-network network, whole object loss function is to set
The sum of letter loss and position loss, expression formula is as follows:
L in formulaconfFor confidence loss, LlocIt is lost for position, here using Smooth L1 Loss；N be with it is preparatory
The matched default frame number of callout box；α is the weight for balancing confidence loss and position loss, is usually arranged as 1；Z is default frame
With the matching result of different classes of preparatory callout box；C is the confidence level for predicting object frame；L is the position letter for predicting object frame
Breath；G is the location information of preparatory callout box.
Further, algorithm is based on deep learning thought and is studied, in the training stage of algorithm, by these default frames and
Preparatory callout box matching, as a result entrance loss function is to train matching strategy.In forecast period, then directly the inclined of frame is defaulted in prediction
It moves and to the corresponding score of each classification, obtains final result finally by non-maxima suppression removal redundancy.
In conclusion the embodiment of the present invention has the advantages that compared with prior art
The present invention identifies this problem for high speed surface unmanned boat detection of obstacles, analyzes and establishes in conjunction with actual conditions
Algorithm model proposes the high speed surface unmanned boat visible detection method based on deep learning algorithm.Based on lightweight
MobileNet convolutional layer designs master network and extracts full figure feature, and adds the characteristic pattern of multireel lamination output different scale, base
Multiple scale detecting is realized in SSD thought design detection sub-network network.And a new target data set waterborne is created, use training set
The training for completing entire detection system on NVIDIA GeForce 1080TI to the network, in NVIDIA Jetson TX2
On verified with test set, the results showed that this method can under to more scenes different size of ship and motor boat etc.
While obstacle target keeps 82.1% average detected precision, the detection time of single frame video is less than 50ms, better than existing
Same type scheme.
In order to explain the structural features and functions of the invention more clearly, come with reference to the accompanying drawing with specific embodiment to this hair
It is bright to be described in detail.
Detailed description of the invention
Fig. 1 is the vision-based detection real time algorithm model of the invention based on deep learning.
Fig. 2 is multiple dimensioned detection of obstacles result:
A) multiple similar targets；B) multiple similar multiscale targets；C) multi-class targets；D) multiscale target.
Fig. 3 is complex scene testing result:
A) spray scene；B) complex building scene；C) extra large day scene；D) pure water face scene.
Fig. 4 is that bloom shines scene detection results:
A) more noise scenarios；B) change colour scene.
Specific embodiment
The following further describes the technical solution of the present invention in the following with reference to the drawings and specific embodiments.
The present invention identifies this problem for high speed surface unmanned boat detection of obstacles, analyzes and establishes in conjunction with actual conditions
Algorithm model proposes the high speed surface unmanned boat visible detection method based on deep learning algorithm.First with lightweight master network
Full figure feature is extracted, then by detection sub-network network fusion multi-characteristic as a result, and screening out redundant results with non-maxima suppression
Regression forecasting location parameter simultaneously, final output testing result.
As shown in Figure 1, the size normalizing of picture is 300 × 300 by input picture, first master network, then to picture square
Battle array carries out feature extraction using each convolutional layer, and the present invention forms convolutional layer using depth convolution sum point convolution, and depth convolution will
Each convolution kernel is applied to each channel, and puts the output that convolution is used to combine channel convolution, in lower network model parameter
It is handled in the case where amount and calculation amount, exports the characteristic pattern of extracted different scale.
Depth separates the computing cost ratio of convolutional layer comparison with standard convolutional layer are as follows:
In formula: DFFor the width and height of input feature vector figure；M is the quantity of input channel；DKFor Standard convolution layer and depth
The convolution kernel size of convolution；N is the quantity of convolution kernel.
Then the various sizes of characteristic pattern of detection sub-network network extraction is used as input to realize multiple scale detecting, this 6 features
The size of figure is respectively 19 × 19,10 × 10,5 × 5,3 × 3,2 × 2,1 × 1.Each characteristic area of the algorithm to input feature vector figure
Domain generates a series of default frame of different sizes, different proportion, and size and ratio are related with corresponding characteristic layer, it is assumed that model
M layers of characteristic pattern are used when detection, then the default frame ratio calculation formula of k-th of characteristic pattern is as follows:
[1, m] k ∈ in formula, wherein SmaxAnd SminRepresent default frame minimum and maximum ratio shared in character pair figure
Example, is respectively set to 0.2 and 0.9.
By providing different the ratio of width to height to the default frame on same characteristic layer to enhance the robustness to body form, fit
Detection case for different size of ship.Setting the ratio of width to height is aspect_ratio=(1,2,1/2,3,1/3), specific wide
It is determined with height by following formula:
Particularly, when aspect_ratio is 1, it is specified that default frame parameter are as follows:
Characteristic pattern makes it also can be corresponding different in the receptive field of picture with the difference of corresponding frame size, forms more rulers
The detection of degree exports testing result as shown in Figure 1.
Target classification and position are returned simultaneously when detection sub-network network training, whole object loss function is confidence damage
Become estranged position loss the sum of, expression formula is as follows:
L in formulaconfFor confidence loss, LlocIt is lost for position, here using Smooth L1 Loss；N be with it is preparatory
The matched default frame number of callout box；α is the weight for balancing confidence loss and position loss, is usually arranged as 1；Z is default frame
With the matching result of different classes of preparatory callout box；C is the confidence level for predicting object frame；L is the position letter for predicting object frame
Breath；G is the location information of preparatory callout box.
Fig. 2, Fig. 3 and Fig. 4 are respectively multiscale target scene, complex scene and bloom according to detection knot of the invention under scene
Fruit, has a look at out, and the present invention has stronger robustness, and multiple dimensioned accurate detection can be kept under various complex environments.
In conclusion the above is merely preferred embodiments of the present invention, being not intended to limit the scope of the present invention.
All within the spirits and principles of the present invention, any modification, equivalent replacement, improvement and so on should be included in of the invention
Within protection scope.
Claims (6)
1. a kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat, total algorithm model can be divided into two parts,
First part based on MobileNet convolutional layer designs lightweight master network to generate the characteristic patterns of multiple and different scales, and second
Divide and detection sub-network network built based on SSD thought, merges each layer characteristic pattern and final output, it is characterised in that:
One, projected depth separates convolutional layer in lightweight master network, and Standard convolution layer is resolved into a depth convolution sum one
A convolution: each convolution kernel is applied to each channel by depth convolution, and point convolution is used to combine the output of channel convolution, deep
Spend the computing cost ratio of separable convolutional layer comparison with standard convolutional layer are as follows:
Two, lightweight master network removes the last average pond layer of MobileNet, full articulamentum and Softmax layers, and in network
8 convolutional layers are added below；
Three, detection sub-network network extracts various sizes of characteristic pattern as input, and raw to each characteristic area of input feature vector figure
At a series of different sizes, the default frame of different proportion, it is assumed that use m layers of characteristic pattern when model inspection, then k-th characteristic pattern
It is as follows to default frame ratio calculation formula:
[1, m] k ∈ in formula, wherein SmaxAnd SminDefault frame minimum and maximum ratio shared in character pair figure is represented, point
0.2 and 0.9 are not set as it；
Characteristic pattern makes it also can be corresponding different in the receptive field of picture with the difference of corresponding frame size, is formed multiple dimensioned
Detection；
Four, target classification and position are returned simultaneously when the training of detection sub-network network, whole object loss function is confidence damage
Become estranged position loss the sum of, expression formula is as follows:
L in formulaconfFor confidence loss, LlocIt is lost for position, here using Smooth L1 Loss；N is and marks in advance
The matched default frame number of frame；α is the weight for balancing confidence loss and position loss, is set as 1；Z is default frame and inhomogeneity
The matching result of other preparatory callout box；C is the confidence level for predicting object frame；L is the location information for predicting object frame；G is pre-
The location information of first callout box.
2. the water surface sensation target fast algorithm of detecting according to claim 1 based on high speed unmanned boat, which is characterized in that
In three, the size of characteristic pattern is respectively 19 × 19,10 × 10,5 × 5,3 × 3,2 × 2,1 × 1.
3. the water surface sensation target fast algorithm of detecting according to claim 2 based on high speed unmanned boat, which is characterized in that
In three, by providing different the ratio of width to height to the default frame on same characteristic layer, setting the ratio of width to height be aspect_ratio=(1,
2,1/2,3,1/3), specific wide and height is determined by following formula:
4. the water surface sensation target fast algorithm of detecting according to claim 3 based on high speed unmanned boat, which is characterized in that
When aspect_ratio is 1, it is specified that default frame parameter are as follows:
5. the water surface sensation target fast algorithm of detecting according to claim 1 based on high speed unmanned boat, which is characterized in that
In the training stage of algorithm, these default frames and preparatory callout box are matched, as a result entrance loss function is to train matching strategy.
6. the water surface sensation target fast algorithm of detecting according to claim 5 based on high speed unmanned boat, which is characterized in that
In forecast period, then directly prediction defaults the offset of frame and to the corresponding score of each classification, presses down finally by non-maximum
System removal redundancy obtains final result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910617212.7A CN110472500A (en) | 2019-07-09 | 2019-07-09 | A kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910617212.7A CN110472500A (en) | 2019-07-09 | 2019-07-09 | A kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110472500A true CN110472500A (en) | 2019-11-19 |
Family
ID=68507471
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910617212.7A Pending CN110472500A (en) | 2019-07-09 | 2019-07-09 | A kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110472500A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111178283A (en) * | 2019-12-31 | 2020-05-19 | 哈尔滨工业大学（深圳） | Unmanned aerial vehicle image-based ground object identification and positioning method for established route |
| CN111259844A (en) * | 2020-01-21 | 2020-06-09 | 陕西师范大学 | Real-time monitoring method for examinees in standardized examination room |
| CN111582095A (en) * | 2020-04-27 | 2020-08-25 | 西安交通大学 | Light-weight rapid detection method for abnormal behaviors of pedestrians |
| CN111666875A (en) * | 2020-06-05 | 2020-09-15 | 浙江工业大学 | Manufacturing worker fatigue state detection method based on machine vision |
| CN111783784A (en) * | 2020-06-30 | 2020-10-16 | 创新奇智(合肥)科技有限公司 | Method and device for detecting building cavity, electronic equipment and storage medium |
| CN112733924A (en) * | 2021-01-04 | 2021-04-30 | 哈尔滨工业大学 | Multi-patch component detection method |
| CN113484864A (en) * | 2021-07-05 | 2021-10-08 | 中国人民解放军国防科技大学 | Unmanned ship-oriented navigation radar and photoelectric pod collaborative environment sensing method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108921218A (en) * | 2018-06-29 | 2018-11-30 | 炬大科技有限公司 | A kind of target object detection method and device |
| CN109117838A (en) * | 2018-08-08 | 2019-01-01 | 哈尔滨工业大学 | Object detection method and device applied to unmanned boat sensory perceptual system |
-
2019
- 2019-07-09 CN CN201910617212.7A patent/CN110472500A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108921218A (en) * | 2018-06-29 | 2018-11-30 | 炬大科技有限公司 | A kind of target object detection method and device |
| CN109117838A (en) * | 2018-08-08 | 2019-01-01 | 哈尔滨工业大学 | Object detection method and device applied to unmanned boat sensory perceptual system |
Non-Patent Citations (2)
| Title |
|---|
| YITING LI ET AL.: "Research on a Surface Defect Detection Algorithm Based on MobileNet-SSD", 《APPLIED SCIENCES》 * |
| 衣世东: "基于深度学习的图像识别算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111178283A (en) * | 2019-12-31 | 2020-05-19 | 哈尔滨工业大学（深圳） | Unmanned aerial vehicle image-based ground object identification and positioning method for established route |
| CN111259844A (en) * | 2020-01-21 | 2020-06-09 | 陕西师范大学 | Real-time monitoring method for examinees in standardized examination room |
| CN111259844B (en) * | 2020-01-21 | 2023-02-24 | 陕西师范大学 | Real-time monitoring method for examinees in standardized examination room |
| CN111582095A (en) * | 2020-04-27 | 2020-08-25 | 西安交通大学 | Light-weight rapid detection method for abnormal behaviors of pedestrians |
| CN111582095B (en) * | 2020-04-27 | 2022-02-01 | 西安交通大学 | Light-weight rapid detection method for abnormal behaviors of pedestrians |
| CN111666875A (en) * | 2020-06-05 | 2020-09-15 | 浙江工业大学 | Manufacturing worker fatigue state detection method based on machine vision |
| CN111783784A (en) * | 2020-06-30 | 2020-10-16 | 创新奇智(合肥)科技有限公司 | Method and device for detecting building cavity, electronic equipment and storage medium |
| CN112733924A (en) * | 2021-01-04 | 2021-04-30 | 哈尔滨工业大学 | Multi-patch component detection method |
| CN113484864A (en) * | 2021-07-05 | 2021-10-08 | 中国人民解放军国防科技大学 | Unmanned ship-oriented navigation radar and photoelectric pod collaborative environment sensing method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110472500A (en) | A kind of water surface sensation target fast algorithm of detecting based on high speed unmanned boat | |
| CN110782481B (en) | Unmanned ship intelligent decision-making method and system | |
| CN111461023B (en) | Method for quadruped robot to automatically follow pilot based on three-dimensional laser radar | |
| Prasad et al. | Challenges in video based object detection in maritime scenario using computer vision | |
| CN105022990B (en) | A kind of waterborne target rapid detection method based on unmanned boat application | |
| CN109271856A (en) | Remote sensing image object detection method based on expansion residual error convolution | |
| Bovcon et al. | WaSR—A water segmentation and refinement maritime obstacle detection network | |
| CN107609522A (en) | A kind of information fusion vehicle detecting system based on laser radar and machine vision | |
| CN105184816A (en) | Visual inspection and water surface target tracking system based on USV and detection tracking method thereof | |
| CN108153334A (en) | No cooperative target formula unmanned helicopter vision is independently maked a return voyage and drop method and system | |
| CN110458160A (en) | A kind of unmanned boat waterborne target recognizer based on depth-compression neural network | |
| CN111723632B (en) | Ship tracking method and system based on twin network | |
| CN105225251A (en) | Over the horizon movement overseas target based on machine vision identifies and locating device and method fast | |
| Zhang et al. | Research on unmanned surface vehicles environment perception based on the fusion of vision and lidar | |
| Zhang et al. | A object detection and tracking method for security in intelligence of unmanned surface vehicles | |
| CN110766721A (en) | Carrier landing cooperative target detection method based on airborne vision | |
| Woo et al. | Obstacle avoidance and target search of an Autonomous Surface Vehicle for 2016 Maritime RobotX challenge | |
| CN116486287A (en) | Target detection method and system based on environment self-adaptive robot vision system | |
| Shi et al. | Obstacle type recognition in visual images via dilated convolutional neural network for unmanned surface vehicles | |
| CN113933828A (en) | Unmanned ship environment self-adaptive multi-scale target detection method and system | |
| Zhou et al. | A real-time algorithm for visual detection of high-speed unmanned surface vehicle based on deep learning | |
| CN116486252A (en) | Intelligent unmanned search and rescue system and search and rescue method based on improved PV-RCNN target detection algorithm | |
| CN110895680A (en) | Unmanned ship water surface target detection method based on regional suggestion network | |
| CN118314770B (en) | Ship waterway navigation risk early warning method and system | |
| Jeon et al. | Visual and Infrared Detection and Ranging (VAIDAR) for Marine Navigational Hazards |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191119 |